keras_relational.py

#!/usr/bin/env python3
 
 
 
from tensorflow.keras.layers import Layer, Reshape
from tensorflow.keras import activations
from tensorflow.keras.activations import sigmoid, tanh
from tensorflow.keras import backend as K
import numpy as np
 
 
class Relations(Layer):
    """
    This is a class which implements Relational Layer into Keras.
    Relational Layer compares every combination of two feature groups with shared weights.
    Use this class as every other Layer in Keras.
    Relevant Paper: https://arxiv.org/abs/1706.01427
    RN(O) = f_phi(sum_phi(g_theta(o_i,o_j)))
    For flexibility reason only the part g(o_i,o_j) is modelled
    f_phi corresponds to a MLP net
    To sum over all permutations please use GlobalAveragePooling1D from keras.
    """
 
    def __init__(self, number_features, hidden_feature_shape=[30, 30, 30, 30], activation=tanh, **kwargs):
        """
        Init class.
        """
 
        
        self.number_featuresnumber_features = number_features
        
        self.number_groupsnumber_groups = 0
        
        self.hidden_feature_shapehidden_feature_shape = hidden_feature_shape
        
        self.activationactivation = activations.get(activation)
        
        self.group_lengroup_len = 0
        
        self.weightvariablesweightvariables = []
        
        self.combinationscombinations = 0
 
        super().__init__(**kwargs)
 
    def build(self, input_shape):
        """
        Build all weights for Relations Layer
        :param input_shape: Input shape of tensor
        :return:  Nothing
        """
        # only accept 2D layers
        assert(len(input_shape) == 3)
 
        self.number_groupsnumber_groups = input_shape[1]
 
        self.group_lengroup_len = input_shape[2]
 
        self.combinationscombinations = np.int32(np.math.factorial(self.number_groupsnumber_groups) / (2 * np.math.factorial(self.number_groupsnumber_groups - 2)))
 
        dense_shape = [2 * self.group_lengroup_len] + self.hidden_feature_shapehidden_feature_shape + [self.number_featuresnumber_features]
 
        for i in range(len(dense_shape[:-1])):
            weights = self.add_weight(name=f'relation_weights_{i}',
                                      shape=list(dense_shape[i:i + 2]), initializer='glorot_uniform', trainable=True)
            bias = self.add_weight(name=f'relation_weights_{i}',
                                   shape=(dense_shape[i + 1],), initializer='zeros', trainable=True)
 
            self.weightvariablesweightvariables.append([weights, bias])
 
        super().build(input_shape)
 
    def call(self, inputs):
        """
        Compute Relational Layer
        :param inputs: input tensor
        :return: output tensor
        """
        input_groups = [inputs[:, i, :] for i in range(self.number_groupsnumber_groups)]
        outputs = []
        for index, group1 in enumerate(input_groups[:-1]):
            for group2 in input_groups[index + 1:]:
                net = K.dot(K.concatenate([group1, group2]), self.weightvariablesweightvariables[0][0])
                net = K.bias_add(net, self.weightvariablesweightvariables[0][1])
                for variables in self.weightvariablesweightvariables[1:]:
                    net = self.activationactivation(net)
                    net = K.dot(net, variables[0])
                    net = K.bias_add(net, variables[1])
                outputs.append(sigmoid(net))
 
        flat_result = K.concatenate(outputs)
        return Reshape((self.combinationscombinations, self.number_featuresnumber_features,))(flat_result)
 
    def compute_output_shape(self, input_shape):
        """
        Compute Output shape
        :return: Output shape
        """
        # only 2D layers
        assert(len(input_shape) == 3)
 
        self.combinationscombinations = np.int32(np.math.factorial(self.number_groupsnumber_groups) / (2 * np.math.factorial(self.number_groupsnumber_groups - 2)))
 
        return (input_shape[0], self.combinationscombinations, self.number_featuresnumber_features)
 
    def get_config(self):
        """
        Config required for saving parameters in keras model.
        """
        config = {
            'number_features': self.number_featuresnumber_features,
            'hidden_feature_shape': self.hidden_feature_shapehidden_feature_shape,
            'activation': activations.serialize(self.activationactivation)
        }
        base_config = super().get_config()
        return dict(list(base_config.items()) + list(config.items()))
 
 
class EnhancedRelations(Layer):
    """
    This is a class which implements Relational Layer into Keras.
    See Class Relations for details.
    EnhanceRelations use an additional input for passing event information to every comparison:
    RN(O) = f_phi(sum_phi(g_theta(o_i,o_j,q)))
    q is fed in as second one dimensional input.
    """
 
    def __init__(self, number_features, hidden_feature_shape=[30, 30, 30, 30], activation=tanh, **kwargs):
        """
        Init class.
        """
 
        
        self.number_featuresnumber_features = number_features
        
        self.number_groupsnumber_groups = 0
        
        self.hidden_feature_shapehidden_feature_shape = hidden_feature_shape
        
        self.activationactivation = activations.get(activation)
        
        self.group_lengroup_len = 0
        
        self.weightvariablesweightvariables = []
        
        self.combinationscombinations = 0
        
        self.question_lenquestion_len = 0
 
        super().__init__(**kwargs)
 
    def build(self, input_shape):
        """
        Build all weights for Relations Layer
        :param input_shape: Input shape of tensor
        :return:  Nothing
        """
        # accept only 2 inputs
        assert(len(input_shape) == 2)
        # first input should be a 2D layers
        assert(len(input_shape[0]) == 3)
        # second input should be a 1D layers
        assert(len(input_shape[1]) == 2)
 
        self.number_groupsnumber_groups = input_shape[0][1]
 
        self.group_lengroup_len = input_shape[0][2]
 
        self.question_lenquestion_len = input_shape[1][1]
 
        self.combinationscombinations = np.int32(np.math.factorial(self.number_groupsnumber_groups) / (2 * np.math.factorial(self.number_groupsnumber_groups - 2)))
 
        dense_shape = [2 * self.group_lengroup_len + self.question_lenquestion_len] + self.hidden_feature_shapehidden_feature_shape + [self.number_featuresnumber_features]
 
        for i in range(len(dense_shape[:-1])):
            weights = self.add_weight(name=f'relation_weights_{i}',
                                      shape=list(dense_shape[i:i + 2]), initializer='glorot_uniform', trainable=True)
            bias = self.add_weight(name=f'relation_weights_{i}',
                                   shape=(dense_shape[i + 1],), initializer='zeros', trainable=True)
 
            self.weightvariablesweightvariables.append([weights, bias])
 
        super().build(input_shape)
 
    def call(self, inputs):
        """
        Compute Relational Layer
        :param inputs: input tensor
        :return: output tensor
        """
        input_groups = [inputs[0][:, i, :] for i in range(self.number_groupsnumber_groups)]
        questions = inputs[1]
        outputs = []
        for index, group1 in enumerate(input_groups[:-1]):
            for group2 in input_groups[index + 1:]:
                net = K.dot(K.concatenate([group1, group2, questions]), self.weightvariablesweightvariables[0][0])
                net = K.bias_add(net, self.weightvariablesweightvariables[0][1])
                for variables in self.weightvariablesweightvariables[1:]:
                    net = self.activationactivation(net)
                    net = K.dot(net, variables[0])
                    net = K.bias_add(net, variables[1])
                outputs.append(sigmoid(net))
 
        flat_result = K.concatenate(outputs)
        return Reshape((self.combinationscombinations, self.number_featuresnumber_features,))(flat_result)
 
    def compute_output_shape(self, input_shape):
        """
        Compute Output shape
        :return: Output shape
        """
        # accept only 2 inputs
        assert(len(input_shape) == 2)
        # first input should be a 2D layers
        assert(len(input_shape[0]) == 3)
        # second input should be a 1D layers
        assert(len(input_shape[1]) == 2)
 
        self.combinationscombinations = np.int32(np.math.factorial(self.number_groupsnumber_groups) / (2 * np.math.factorial(self.number_groupsnumber_groups - 2)))
 
        return (input_shape[0][0], self.combinationscombinations, self.number_featuresnumber_features)
 
    def get_config(self):
        """
        Config required for saving parameters in keras model.
        """
        config = {
            'number_features': self.number_featuresnumber_features,
            'hidden_feature_shape': self.hidden_feature_shapehidden_feature_shape,
            'activation': activations.serialize(self.activationactivation)
        }
        base_config = super().get_config()
        return dict(list(base_config.items()) + list(config.items()))